TSMC Leads in Switch CPO, Samsung Bets on XPU Optical Packaging
0xBroomberg
TSMC has moved switch-level CPO into customer deployment through live projects with Broadcom and Nvidia, while Samsung is betting on the next stage — optical I/O integrated into XPU compute packaging. The core contest is how close light gets to the chip.
How far along is switch CPO?
Broadcom's 102.4 Tbps CPO Ethernet switch, built on TSMC's COUPE platform, has shipped samples to early customers. Nvidia's Quantum-X photonic switch is already shipping.
This means → CPO — co-packaged optics, mounting the optical engine right next to the switch chip instead of using plug-in modules — has crossed from lab validation into real orders.
First adopters include CoreWeave, Lambda, and Oracle. TSMC's silicon photonics and SoIC — a 3D chip-stacking technology — form the shared manufacturing foundation for this generation.
What exactly are TSMC and Samsung competing over?
TSMC's bet is switch CPO: the optical engine sits beside the switch ASIC. The key challenge is bonding photonic and electronic dies together and integrating them into the switch package. HBM — high-bandwidth memory — is not required at this stage; TSMC's moat rests on mature process and packaging.
Samsung is targeting a different architecture — XPU optical packaging: placing compute chips (XPUs), HBM, and optical engines together on an interposer, making optical I/O part of the compute package itself.
In plain terms = TSMC is solving "how does the switch use light"; Samsung wants to solve "how does the compute chip use light." The latter is harder, but if it works, light moves one step closer to the processing core.
Why move the optical engine closer to the chip?
Samsung's OECC 2026 presentation gives a clear answer in energy numbers: pluggable optics at board level consume roughly 10 picojoules per bit; moving the engine to the substrate near the switch cuts that to about 5 pJ; pushing it onto the interposer next to the XPU drops it to roughly 2 pJ.
This means → each step closer to the compute die cuts energy consumption by roughly half. Power is the core force driving this migration.
What is Samsung's differentiating hand?
Samsung's strongest card is its "trinity" — it makes HBM, foundry logic, and silicon photonics under one roof. TSMC has leading-edge logic, photonics, and CoWoS packaging, but it does not produce HBM.
Samsung EVP Won-Kyoung Choi said at Nano Korea on July 9 that the company is developing 2.xD advanced packaging to integrate HBM, logic, and silicon-photonic dies in a single package.
In plain terms = Samsung's pitch is "we build every critical component in-house," which in theory enables joint co-design of interfaces, I/O, optics, and thermal management — something TSMC cannot do without an external HBM partner.
What is the biggest risk on this path?
2.xD packaging faces extreme multi-die yield pressure: logic, HBM, photonic ICs, electronic ICs, and the interposer all sit in one package. A single component failure scraps the entire assembly.
This means → more dies and higher bonding complexity are multiplying both yield risk and cost. Whether Samsung's "trinity" advantage can survive the manufacturing yield gauntlet is the make-or-break question for this roadmap.
What signal should investors watch next?
The competitive landscape is not standing still. TSMC is integrating COUPE with CoWoS packaging and tapping external HBM through its ecosystem. SK Hynix is investing $3.87 billion in an advanced packaging plant in Indiana, set for volume production in 2028, and has folded CPO into its memory-system R&D roadmap.
Samsung's published "one-stop CPO" roadmap targets 2029. Measured against current switch-CPO shipment volumes and customer qualifications, Samsung has not yet matched TSMC's commercialization cadence.
This reflects a single signal worth tracking over the next 12 months: whether a named customer places a design order that explicitly binds HBM, logic, and optical I/O into one package and awards it to Samsung's foundry — the definitive proof point that Samsung's trinity has converted from a paper asset into a commercial moat.
Content is for reference only, not financial advice.